The principle that we always know more than we can tell, which historically grounded the limit of automating complex cognitive tasks.
Polanyi's paradox is an epistemological principle formulated by Michael Polanyi in his work on tacit knowledge, synthesized in The Tacit Dimension (1966). Its original statement is that we always know more than we can tell: a large share of human competence, from the surgeon's dexterity to the judge's discernment, rests on embodied, contextual and hard-to-formalize knowledge that cannot be fully described by explicit rules. In the automation field, this principle was long interpreted as a structural barrier to the algorithmic substitution of skilled workers: if a task cannot be entirely decomposed into explicit instructions, it cannot be delegated to a machine. David Autor and co-authors formalized this intuition in the ALM model (2003), which anticipated that routine tasks would be automated but that complex cognitive tasks would remain the preserve of humans. The emergence of large language models and agentic AI circumvents this paradox not by resolving the philosophical problem but by producing expert-quality reasoning through statistical pathways, without needing explicit instructions. This empirical circumvention constitutes the central rupture theorized in current research on the economic impact of AI on cognitive labor markets.
An experienced marine insurance underwriter is unable to fully explain why he declines a risk whose objective data look acceptable: he draws on an intuition shaped by twenty years of experience. This is precisely the type of tacit judgment that Polanyi's paradox identified as non-automatable and that foundation models have partially learned to emulate through probabilistic approaches.
Polanyi paradox, savoir tacite, tacit knowledge, paradoxe cognitif